IT之家 9 月 24 日消息,阿里巴巴发布第 4 个癌症筛查 AI 模型。阿里达摩院联合四川省肿瘤医院、中山大学肿瘤防治中心等机构研发食管癌筛查 AI 模型 DAMO EAGLE,无需插管和造影,从平扫 CT 上就能识别食管癌,涵盖早期和癌前恶性病变,已在 3 个国家 8 万多病例上取得验证。
相关论文于 9 月 22 日登上国际顶级期刊《自然 · 医学》(Nature Medicine)。
阿里达摩院介绍称,中国食管癌发病率和死亡率约占全球一半,多数患者直到晚期才发现,失去根治机会。如能早发现,大部分患者可借助内镜下微创手术根治,5 年生存率可达 95% 以上。当下食管癌的标准筛查手段是上消化道内镜,具有一定的侵入性,且操作复杂,难以大范围推广。
胸部平扫 CT 成本低、接受度高,广泛应用于各类医疗情形,尤其是低剂量的平扫 CT,已是肺癌的建议筛查方法。这些 CT 的扫描范围本身就覆盖了食管,筛查肺癌时如能同时识别食管癌风险,无疑是“一举多得”。不过,食管是一个狭长、易塌陷的空腔器官,早期病变隐匿于上皮层和黏膜,再加上心脏及大血管运动影响,影像科医生难以从平扫 CT 中识别。
论文提到,DAMO EAGLE 模型正在攻克这个曾被认为“不可能”的任务。达摩院算法专家姚佳文介绍,研究团队依托多年来在医疗影像 AI 上的工艺积累,并创新地将内镜检查报告和增强 CT 影像上的准确病灶位置配对到平扫 CT 影像上,成功训练 AI 模型识别人眼难以分辨的早期食管癌病灶。
论文表明,在机会性筛查情形下,DAMO EAGLE 模型对食管癌的敏感性(不漏报的能力)达到 90%,针对信号颇为微弱的癌前病变,敏感性也有 52.5%。模型在真实情形中的特异性(不误报的能力)高达 99.2%。值得一提的是,模型在低剂量 CT 上表现未见下降,有望无缝切入现有的肺癌筛查流程中,不增加患者负担。
四川省肿瘤医院放疗科王奇峰主任医师介绍,受热烫和高盐饮食习惯影响,中国西部地区特别是四川省食管癌较为高发,受限于医疗资源和依从性,内镜筛查普及尚有较大缺口。他表示:“AI 不是为了替代内镜,而是筛选出高风险个体再开展内镜检查,这样可以让内镜检查更精准,大幅提高检出率。更关键的是,AI 能将防治关口前移,额外给出一次早发现的机会。”
IT之家获悉,随着胰腺癌、胃癌、肠癌、食管癌、主动脉夹层等 AI 模型发布,达摩院已跑通“平扫 CT+AI”一扫多查原创工艺路线,发表《自然 · 医学》论文 5 篇,有望借助一次平扫 CT 检查出中国死亡人数最多的前七大癌症。
本文来源:https://www.ithome.com/1/006/720.htm
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。
Thanks to ITHome users "Buyao Ya" and "Different Feelings" for their leads!
ITHome, September 24 news: Alibaba has released its 4th cancer screening AI model. DAMO Academy at Alibaba, in collaboration with institutions including the Sichuan Cancer Hospital and the Sun Yat-sen University Cancer Center, has developed the esophageal cancer screening AI model, DAMO EAGLE. It does not require intubation or contrast agents and can identify esophageal cancer from non-contrast CT scans, covering both early-stage and premalignant lesions. The model has been validated on over 80,000 cases across 3 countries.
The related paper was published on September 22 in the top international journal "Nature Medicine".
DAMO Academy at Alibaba notes that China accounts for approximately half of the global incidence and mortality rate of esophageal cancer, with most patients not diagnosed until late stages, losing the chance for curative treatment. If detected early, the majority of patients can be cured through minimally invasive endoscopic surgery, with a 5-year survival rate exceeding 95%. Currently, the standard screening method for esophageal cancer is upper gastrointestinal endoscopy, which is invasive, complex, and difficult to scale.
Chest non-contrast CT is low-cost, has high acceptance, and is widely used in various medical scenarios. Especially low-dose non-contrast CT, which is a recommended screening method for lung cancer, inherently covers the esophagus in its scan range. Identifying esophageal cancer risk while screening for lung cancer would be a "win-win". However, the esophagus is a narrow, prone-to-collapse hollow organ; early lesions are hidden in the epithelial and mucosal layers, and affected by the movement of the heart and major blood vessels, making it difficult for radiologists to identify them from non-contrast CTs.
The paper states that the DAMO EAGLE model is tackling a task previously considered "impossible". Yao Jiawen, an algorithm expert at DAMO Academy, introduced that the research team relied on years of accumulation in medical imaging AI and innovatively paired accurate lesion locations from endoscopic examination reports and enhanced CT images onto non-contrast CT images. This successfully trained the AI model to identify early esophageal cancer lesions that are difficult for the human eye to distinguish.
The paper shows that in opportunistic screening scenarios, the DAMO EAGLE model achieves a 90% sensitivity (ability to avoid missed diagnoses) for esophageal cancer, and a 52.5% sensitivity for faint premalignant lesions. The model's specificity (ability to avoid false positives) in real-world scenarios is as high as 99.2%. Notably, the model's performance does not decrease on low-dose CTs, promising a seamless integration into existing lung cancer screening workflows without adding burden to patients.
Dr. Wang Qifeng, Chief Physician of the Radiotherapy Department at Sichuan Cancer Hospital, notes that due to dietary habits of eating very hot and high-salt food, esophageal cancer is highly prevalent in western China, particularly in Sichuan Province. Due to limited medical resources and compliance issues, there is a significant gap in the widespread adoption of endoscopic screening. He stated: "AI is not intended to replace endoscopy, but to screen high-risk individuals for endoscopic examinations. This makes endoscopic examinations more precise and significantly increases the detection rate. More importantly, AI can shift prevention and control earlier, providing an additional chance for early detection."
ITHome has learned that with the release of AI models for pancreatic cancer, gastric cancer, colorectal cancer, esophageal cancer, and aortic dissection, DAMO Academy has established an original "non-contrast CT + AI" technology route for multiple diagnoses in a single scan. It has published 5 papers in "Nature Medicine" and is expected to use a single non-contrast CT to detect China's top seven deadliest cancers.


